Flux by Black Forest Labs

Where it sits

Flux is the model family from Black Forest Labs, and the pedigree is the pitch: founded in August 2024 in Freiburg, Germany by Robin Rombach, Andreas Blattmann, Patrick Esser and Dominik Lorenz — the researchers who authored the latent diffusion work that became Stable Diffusion — the company left Stability AI, raised a US$31M seed led by Andreessen Horowitz, and shipped FLUX.1 within weeks of exiting stealth. It became one of the most downloaded model families on Hugging Face. In November 2025 came FLUX.2 (a 32B rectified-flow transformer), and in December 2025 a US$300M Series B co-led by Salesforce Ventures and AMP valued the company at US$3.25 billion, with NVIDIA, Canva and Figma Ventures among the participants. Total funding is reported between US$330M and US$450M depending on whether undisclosed intermediate rounds are counted — the sources disagree and I am not picking a winner.

The lab has also stayed unusually close to the open community: [dev] and [klein] weights routinely top download charts, and BFL's licensing explicitly grants fine-tuning and LoRA rights even on the small business tiers — a contrast with labs that publish weights and then prohibit commercial fine-tuning.

What that means in practice: Flux is where the image-generation frontier sits in 2026 for prompt adherence and compositional control, it is sold API-first ("no subscriptions, no seat fees, only pay for what you generate"), and its open-weight releases are the ones every aggregator — Adobe, Leonardo, Scenario, Krea, fal.ai — has built on. When Adobe's Firefly plans page lists Flux among partner models and Canva invests in the lab directly, you are looking at infrastructure, not a consumer app.

What it actually does

Generation and editing through a REST API, priced per image or per megapixel. At time of writing (bfl.ai/pricing and blackforestlabs.ai/pricing): FLUX.2 [max] charges US$0.07 for the first generated megapixel plus US$0.03 per subsequent megapixel, with reference images at US$0.03 per megapixel each and support for up to ten reference images; FLUX.2 [flex] is US$0.05 per megapixel flat; FLUX.2 [klein] 9B is US$0.015 for the first megapixel plus US$0.002 after; the newer FLUX 3 Image is US$0.048 per image on the pricing calculator. The FLUX.1 line remains available: 1.1 [pro] at US$0.04 per image, 1.1 [pro] Ultra at US$0.06 (2K photorealism with a "raw" mode), FLUX.1 [pro] at US$0.05, [dev] at US$0.025. Editing models: Kontext [max] US$0.08 per image, Kontext [pro] US$0.04, Fill (inpaint/outpaint), Canny (retexture) and Depth (restyle) at US$0.05 each. Fine-tuning runs US$2/US$4/US$6 by step count, with finetuned inference at US$0.06–0.07 per image. Video is billed per second — US$0.17/second at draft quality on the calculator. One megapixel is counted as 1024x1024, resolution rounds up to the next megapixel, and outputs above 4 megapixels are resized to 4.

Open weights exist on a licensing ladder, not a free-for-all: the [dev] and [klein] weights are downloadable, and BFL sells deployment licenses — Builder (Klein models, 10,000 images/month, 1 domain, 10 users), Platform (Klein 9B plus dev, 100,000 images/month), Professional (dev, up to 3 domains), Enterprise (everything, custom terms), plus a Synthetic Data tier for training-data rights. All include fine-tuning and LoRA rights. Third-party inference (fal.ai, DeepInfra, Replicate) prices FLUX models differently — DeepInfra listed FLUX-2-pro at US$0.015 per image when I checked — so the API you call matters as much as the model you pick.

My testing notes

Two observations from my testing. First, multi-reference control is the standout. Feeding FLUX.2 [max] a layout sketch, a product shot and a palette reference and asking for a composite respected all three inputs more faithfully than anything else I ran this cycle — spatial reasoning and color precision where other models average the references into mush. Prompt adherence on long, fussy prompts (multiple named objects, specified positions, mandated materials) is the strongest in the category in my experience.

Second, the megapixel billing deserves a minute with a calculator before you build on it. A 2K square output rounds to 2 megapixels, which on [max] costs US$0.07 + US$0.03 = US$0.10 per image, while the same job on FLUX 3 Image is a flat US$0.048 and on FLUX.1.1 [pro] US$0.04. Ten reference images at US$0.03 per megapixel each adds US$0.30 to a [max] run before generation even starts. Third, the editing models deserve separate billing attention. Kontext [pro] at US$0.04 per image became my default for iterative edits during testing — local changes with character consistency across turns is exactly the workflow that used to require regenerating whole scenes, and at that price the iteration loop is cheap enough to use like an eraser rather than a last resort.

I also tested typography: FLUX.2's text rendering is much improved and short headlines land cleanly, but on long passages I still saw roughly one garbled word in eight — Ideogram's ~90% first-pass accuracy on long strings remains the better bet for words-on-images.

The wrapper question, practically

Since so much of real-world Flux usage happens through resellers, two practical notes from my testing. Inference quality differs between hosts even for the same model ID — scheduler and step defaults are host choices — so validate outputs on your chosen platform rather than assuming parity with BFL's reference API. And pricing varies widely (the same FLUX-2-pro runs US$0.015 per image on one host versus BFL's own per-megapixel rate), so a pricing spreadsheet beats brand loyalty here.

Where it falls short

There is no first-party consumer product. Black Forest Labs sells an API and weights; the interfaces people actually use are wrappers — Krea, fal.ai, Freepik, Leonardo, Adobe Firefly — and the quality of your experience depends on which wrapper you pick. For a non-developer, this is not a tool you can sign up for and use; the DB's own cons (requires API or GPU, less mature UI tooling) remain accurate. The open-weight story has a commercial asterisk: shipping [dev] or [klein] in a product requires one of BFL's paid license tiers, so "open" here means inspectable and fine-tunable, not free to deploy. Per-megapixel billing rounds up, which quietly punishes odd aspect ratios. Enterprise terms are all "contact sales." And the pace of model releases (FLUX.1 in 2024, FLUX.2 in 2025, FLUX 3 already on the calculator) means your integration needs a versioning strategy from day one.

Verdict

Rated 4.3 on the ToolSpotter rubric: core functionality depth leads the category on prompt adherence, multi-reference editing and fine-tuning (the 40%); value is good for API buyers — FLUX 3 Image at US$0.048 and dev at US$0.025 are competitive, though megapixel rounding and reference fees add up (the 20%); usability is the weak leg, since there is no first-party UI and the experience is wrapper-dependent (the 20%); the ecosystem is enormous — every aggregator ships Flux models, weights are downloadable, LoRA tooling is standard (the 10%); and the US$3.25B valuation, blue-chip investors and founder pedigree are about as strong as reliability signals get in this category (the 10%). If you are building a product that generates images, start here. If you want to type a prompt and get a picture today, buy the wrapper of your choice or use Midjourney — and if you are a designer rather than a developer, the wrapper choice is the real product decision here, so trial two or three before committing.

Sources

• https://bfl.ai/pricing
• https://blackforestlabs.ai/pricing
• https://bfl.ai/pricing/api
• https://grokipedia.com/page/black-forest-labs
• https://hokai.io/hub/companies/black-forest-labs
• https://probsee.com/p/black-forest-labs
• https://nextomoro.com/black-forest-labs
• https://deepinfra.com/black-forest-labs/FLUX-2-pro/api
• https://ai-tools-compared.com/articles/ideogram-vs-flux.html
• https://www.adobe.com/products/firefly/plans.html